> The liquid-cooled AI racks being used today for training and inference workloads almost certainly have far higher power output than that.

75kW is a sane "default baseline" and you can find plenty of deployments at 130kW.

There's talk of pushing to 240kW and beyond...